Nvidia research published this week argues that the agent harness — the scaffolding around the model — is doing more of the work than the model itself. Per TechCrunch, the finding is that agents can perform well on a task, and avoid going off the deep end, through fine-tuning, even when the underlying model isn't especially good at that task to begin with.
That cuts against the default assumption that agent quality is downstream of frontier model quality — and it is a convenient argument for a company that sells the compute you'd use to fine-tune. Worth reading the write-up for the specifics before drawing conclusions about your own stack.
Sources: TechCrunch
